A REMARK ON A CENTRAL LIMIT THEOREM FOR …Math. J. Okayama Univ. 60 (2018), 109{135 A REMARK ON A...
Transcript of A REMARK ON A CENTRAL LIMIT THEOREM FOR …Math. J. Okayama Univ. 60 (2018), 109{135 A REMARK ON A...
Math. J. Okayama Univ. 60 (2018), 109–135
A REMARK ON A CENTRAL LIMIT THEOREM FOR
NON-SYMMETRIC RANDOM WALKS ON CRYSTAL
LATTICES
Ryuya Namba
Abstract. Recently, Ishiwata, Kawabi and Kotani [4] proved two kindsof central limit theorems for non-symmetric random walks on crystallattices from the view point of discrete geometric analysis developedby Kotani and Sunada. In the present paper, we establish yet anotherkind of the central limit theorem for them. Our argument is based on ameasure-change technique due to Alexopoulos [1].
1. Introduction and results
Let X = (V,E) be a locally finite, connected and oriented graph. HereV is the set of all vertices and E the set of all oriented edges. For anedge e ∈ E, we denote by o(e), t(e) and e the origin, the terminus andthe inverse edge of e, respectively. We denote by Ex the collection of alledges whose origin is x ∈ V . A path c in X with length n is a sequencec = (e1, . . . , en) of edges ei with t(ei) = o(ei+1) (i = 1, . . . , n−1). We denoteby Ωx,n(X) (x ∈ V, n ∈ N ∪ ∞) the set of all paths of length n for whichorigin o(c) = x. We also denote by o(c), t(c) the origin and the terminus ofthe path c. For simplicity, we write Ωx(X) := Ωx,∞(X).
Let p : E −→ (0, 1] be a transition probability, that is, a positive functionon E satisfying
(1.1)∑e∈Ex
p(e) = 1 (x ∈ V ).
This induces the probability measure Px on Ωx(X) and the random walk as-sociated with the transition probability p is the time homogeneous Markovchain (Ωx(X),Px, wn∞n=0) with values in X defined by wn(c) := o(en+1)(c = (e1, e2, . . . ) ∈ Ωx(X)
). The graph X is endowed with the graph dis-
tance. For a topological space T , we denote by C∞(T ) the space of allfunctions f : T −→ R vanishing at infinity with the uniform topology ∥f∥T∞.
The (p-)transition operator L acting on C∞(X) is defined by
Lf(x) :=∑e∈Ex
p(e)f(t(e)
)(x ∈ V ).
Mathematics Subject Classification. Primary 60J10; Secondary 60F05.Key words and phrases. crystal lattice, central limit theorem, non-symmetric random
walk, (modified) harmonic realization.
109
110 R. NAMBA
The n-step transition probability p(n, x, y) (n ∈ N, x, y ∈ V ) is given byp(n, x, y) := Lnδy(x), where δy(·) stands for the Dirac delta function withpole at y. If there is a positive function m : V −→ (0,∞) up to constantmultiple such that
p(e)m(o(e)
)= p(e)m
(t(e)
)(e ∈ E),
the random walk is called (m-)symmetric or reversible. Otherwise, it iscalled (m-)non-symmetric.
As one of the most fundamental examples of infinite graphs, a crystal lat-tice has been studied by many authors from both geometric and probabilisticviewpoints. Roughly speaking, an infinite graph X is called a (Γ-)crystallattice if X is an infinite-fold covering graph of a finite graph whose coveringtransformation group Γ is abelian. Typical examples we have in mind arethe square lattice, the triangular lattice and the hexagonal lattice, and soon. For basic results, see [5, 6, 7, 8] and literatures therein.
In the theory of random walks on infinite graphs, to investigate the longtime asymptotics, for instance, the central limit theorem (CLT) is a princi-pal topic for both geometers and probabilists. Recently, Ishiwata, Kawabiand Kotani [4] proved two kinds of functional CLTs for non-symmetric ran-dom walks on crystal lattices using the theory of discrete geometric analysisdeveloped by Kotani and Sunada. For more details on discrete geometricanalysis, see Section 2. We also refer to [9, 7].
Before stating our results, we start with a brief review of the setting andresults in [4]. Let us consider a (Γ-)crystal lattice X = (V,E), where thecovering transformation group Γ, acting on X freely, is a torsion free, finitelygenerated abelian group. Here we may assume that Γ is isomorphic to Zd,without loss of generality. We denote by X0 = (V0, E0) its (finite) quotientgraph Γ\X. Let p : E −→ (0, 1] be a Γ-invariant transition probability.Namely, it satisfies (1.1) and p(σe) = p(e) for every σ ∈ Γ, e ∈ E. Throughthe covering map π : X −→ X0, the transition probability p also induces aMarkov chain with values in X0. Since the transition probability p on X ispositive, the random walk on X is irreducible, that is, for every x, y ∈ V ,there exists n = n(x, y) ∈ N such that p(n, x, y) > 0. And so is the randomwalk on X0. Then, thanks to the Perron–Frobenius theorem, we find aunique invariant probability measure on V0 with
(1.2)∑x∈V0
m(x) = 1 and m(x) =∑
e∈(E0)x
p(e)m(t(e)
)(x ∈ V0).
It is also called a stationary distribution (see e.g., Durrett [2]). We also writem : V −→ (0, 1] for the Γ-invariant lift of m to X.
Let H1(X0,R) and H1(X0,R) be the first homology group and the first co-homology group ofX0, respectively. We take a linear map ρR from H1(X0,R)
A CENTRAL LIMIT THEOREM ON CRYSTAL LATTICES 111
onto Γ ⊗ R through the covering map π : X −→ X0. We define the homo-logical direction of the random walk on X0 by
γp :=∑e∈E0
p(e)m(o(e)
)e ∈ H1(X0,R),
and we call ρR(γp)(∈ Γ⊗ R) the asymptotic direction. We remark that therandom walk is m-symmetric if and only if γp = 0. Moreover, γp = 0 impliesρR(γp) = 0. However, the converse does not hold in general. We write g0 forthe (p-)Albanese metric on Γ⊗R. (See Section 2, for its precise definition.)We call that a periodic realization Φ0 : X −→ Γ⊗R is (p-)modified harmonicif
(1.3)∑e∈Ex
p(e)(Φ0
(t(e)
)− Φ0
(o(e)
))= ρR(γp) (x ∈ V ).
This notion was first proposed in [7] to seek the most canonical periodicrealization of a topological crystal in the geometric context.
Now we are in a position to review two kinds of CLTs formulated in [4].We first set a reference point x∗ ∈ V with Φ0(x∗) = 0 and put ξn(c) :=Φ0
(wn(c)
) (n = 0, 1, 2, . . . , c ∈ Ωx∗(X)
). Let C0
([0,∞), (Γ ⊗ R, g0)
)be
the set of all continuous paths from [0,∞) to (Γ ⊗ R, g0) starting from theorigin. We equip it with the usual compact uniform topology. We define ameasurable map X(n) : Ωx∗(X) −→
(C0
([0,∞), (Γ⊗ R, g0)
), µ)by
X(n)t (c) :=
1√n
ξ[nt](c)+(nt−[nt])
(ξ[nt]+1(c)−ξ[nt](c)
)−ntρR(γp)
(t ≥ 0),
where µ is the Wiener measure on the path space C0
([0,∞), (Γ ⊗ R, g0)
).
We write P(n) (n = 1, 2, . . . ) for the probability measure on C0
([0,∞), (Γ⊗
R, g0))induced by X(n). Then the CLT of the first kind is stated as follows:
Theorem 1.1. ([4, Theorem 2.2]) The sequence P(n)∞n=1 of probabilitymeasures converges weakly to the Wiener measure µ as n → ∞. In otherwords, the sequence X(n)∞n=1 converges to a (Γ ⊗ R, g0)-valued standardBrownian motion (Bt)t≥0 starting from the origin in law.
Next we introduce a family pε0≤ε≤1 of transition probabilities on X bypε(e) := p0(e) + εq(e) (e ∈ E), where
p0(e) :=1
2
(p(e) +
m(t(e)
)m(o(e)
)p(e)),q(e) :=
1
2
(p(e)−
m(t(e)
)m(o(e)
)p(e)) (e ∈ E).
112 R. NAMBA
This is nothing but the interpolation of the original transition probabilityp = p1 and the m-symmetric transition probability p0. We should note that,in this setting, the relation ρR(γpε) = ερR(γp) plays a crucial role to obtain
the CLT of the second kind. We write g(ε)0 for the (pε-)Albanese metric on
Γ⊗R. Moreover, let Φ(ε)0 : X −→ (Γ⊗R, g(0)0 ) be the pε-modified harmonic
realization of X.Now set a reference point x∗ ∈ V satisfying Φ
(ε)0 (x∗) = 0 for all 0 ≤ ε ≤ 1
and put ξ(ε)n (c) := Φ
(ε)0
(wn(c)
) (n = 0, 1, 2, . . . , c ∈ Ωx∗(X)
). We define a
measurable map Y(ε,n) : Ωx∗(X) −→ C0
([0,∞), (Γ⊗ R, g(0)0 )
)by
Y(ε,n)t (c) :=
1√n
ξ(ε)[nt](c) + (nt− [nt])
(ξ(ε)[nt]+1(c)− ξ
(ε)[nt](c)
)(t ≥ 0).
Let ν be the probability measure on C0
([0,∞), (Γ ⊗ R, g(0)0 )
)induced by
the stochastic process(Bt + ρR(γp)t
)t≥0
, where (Bt)t≥0 is a (Γ ⊗ R, g(0)0 )-
valued standard Brownian motion with B0 = 0. Moreover let Q(ε,n) be the
probability measure on C0
([0,∞), (Γ ⊗ R, g(0)0 )
)induced by Y(ε,n). Then
the CLT of the second kind is the following.
Theorem 1.2. ([4, Theorem 2.4]) The sequence Q(n−1/2,n)∞n=1 of proba-bility measures converges weakly to the probability measure ν as n → ∞. In
other words, the sequence Y(n−1/2,n)∞n=1 converges to a (Γ⊗R, g(0)0 )-valuedstandard Brownian motion with drift ρR(γp) starting from the origin in law.
The main purpose of the present paper is to establish yet another CLT fornon-symmetric random walks on crystal lattices. Our approach is inspiredby a measure-change techinique due to Alexopoulos [1] in which several limittheorems for random walks on discrete groups of polynomial volume growthare obtained.
Now consider the (m-)non-symmetric transition probability p : E −→(0, 1]. In particular, we assume that ρR(γp) = 0. Let Φ0 : X −→ (Γ⊗R, g0)be a (p-)modified harmonic realization of X. We define a function F =Fx(λ) : V0 ×Hom(Γ,R) −→ (0,∞) by
(1.4) Fx(λ) :=∑
e∈(E0)x
p(e) exp(Hom(Γ,R)
⟨λ,Φ0
(t(e)
)− Φ0
(o(e)
)⟩Γ⊗R
)for x ∈ V0, λ ∈ Hom(Γ,R), where e stands for a lift of e ∈ E0 to X. Then wecan verify that, for every x ∈ V0, the function Fx(·) : Hom(Γ,R) −→ (0,∞)has a unique minimizer λ∗ = λ∗(x) ∈ Hom(Γ,R). (See Lemma 3.1.) We
A CENTRAL LIMIT THEOREM ON CRYSTAL LATTICES 113
define a positive function p : E0 −→ (0, 1] by
(1.5) p(e) :=p(e) exp
(Hom(Γ,R)
⟨λ∗(o(e)
),Φ0
(t(e)
)− Φ0
(o(e)
)⟩Γ⊗R
)Fo(e)
(λ∗(o(e))
)for e ∈ E0. Then it is straightforward to check that the function p also
gives a transition probability on X0. Noting the random walk w(p)n ∞n=0
associated with p is also irreducible, the Perron–Frobenius theorem yields aunique normalized invariant measure m : V0 −→ (0, 1] in the sense of (1.2).We write p : E −→ (0, 1] and m : V −→ (0, 1] for the Γ-invariant lifts of
p : E0 −→ (0, 1] and m : V0 −→ (0, 1], respectively. We write g(p)0 for the
(p-)Albanese metric associated with the transition probability p.Let L(p) be the transition operator, acting on C∞(X), associated with the
transition probability p. Namely,
L(p)f(x) =∑e∈Ex
p(e)f(t(e)
)(x ∈ V ).
Recalling that the function F = Fx(λ) has the (unique) minimizer λ∗ =λ∗(x) for every x ∈ V0, it follows that
(1.6) L(p)Φ0(x)−Φ0(x) =∑e∈Ex
p(e)(Φ0
(t(e)
)−Φ0
(o(e)
))= 0 (x ∈ V ).
This equation means that the p-modified harmonic realization Φ0 : X −→Γ ⊗ R is a p-harmonic realization. In particular, we obtain ρR(γp) = 0.Here we should emphasize that the transition probability p : E0 −→ (0, 1]coincides with the original one p : E0 −→ (0, 1] provided that ρR(γp) = 0.
We fix a reference point x∗ ∈ V such that Φ0(x∗) = 0 and put
ξ(p)n (c) := Φ0
(w(p)n (c)
) (n = 0, 1, 2, . . . , c ∈ Ωx∗(X)
).
We define a measurable map X(n) : Ωx∗(X) −→(C0
([0,∞), (Γ⊗R, g(p)0 )
), µ)
by
(1.7) X(n)t (c) :=
1√n
ξ(p)[nt](c)+(nt− [nt])
(ξ(p)[nt]+1(c)− ξ
(p)[nt](c)
)(t ≥ 0),
where µ = µ(p) is the Wiener measure on C0
([0,∞), (Γ⊗R, g(p)0 )
). We write
P(n) (n = 1, 2, . . . ) for the probability measure on C0
([0,∞), (Γ ⊗ R, g(p)0 )
)induced by X(n). Then our main theorem is stated as follows:
Theorem 1.3. The sequence P(n)∞n=1 of probability measures converges
weakly to the Wiener measure µ as n → ∞. Namely, the sequence X(n)∞n=1
converges to a (Γ⊗R, g(p)0 )-valued standard Brownian motion (B(p)t )t≥0 start-
ing from the origin in law.
114 R. NAMBA
Finally, we give a relationship between the n-step transition probabilitiesp(n, x, y) and p(n, x, y) as follows:
Theorem 1.4. There exist some positive constants C1 and C2 such that
C1p(n, x, y) exp(nMp
)≤ p(n, x, y) ≤ C2p(n, x, y) exp
(nMp
)for all n ∈ N and x, y ∈ V , where
Mp :=∑x∈V0
m(x)(Hom(Γ,R)
⟨λ∗(x), ρR(γp)
⟩Γ⊗R − logFx
(λ∗(x)
)).
For the precise long time asymptotic behavior of p(n, x, y), we refer to [4,Theorem 2.5] and Trojan [10].
The rest of the present paper is organized as follows: In Section 2, weprovide a brief review on the theory of discrete geometric analysis. In Sec-tion 3, we state our measure-change technique in details and give proofs ofTheorems 1.3 and 1.4. We also discuss a relationship between our measure-change technique and a discrete analogue of Girsanov’s theorem due to Fujita[3]. Finally, in Section 4, we give some concrete examples of non-symmetricrandom walks on crystal lattices.
2. A quick review on discrete geometric analysis
In this section, we give basic materials of the theory of discrete geometricanalysis on graphs quickly. For more details, we refer to Kotani–Sunada [7]and Sunada [9].
We consider a random walk on a finite graph X0 = (V0, E0) associatedwith a transition probability p : E0 −→ (0, 1]. Thanks to the Perron–Frobenius theorem, there is a unique (normalized) invariant measure m :V0 −→ (0, 1] in the sense of (1.2). The random walk is called (m-)symmetricif p(e)m
(o(e)
)= p(e)m
(t(e)
)(e ∈ E0).
First we define the 0-chain group and the 1-chain group of X0 by
C0(X0,R) := ∑
x∈V0
axx∣∣∣ ax ∈ R
,
C1(X0,R) := ∑
e∈E0
aee∣∣∣ ae ∈ R, e = −e
,
respectively. Let ∂ : C1(X0,R) −→ C0(X0,R) be the boundary map, givenby the homomorphism satisfying ∂(e) := t(e) − o(e) (e ∈ E0). The firsthomology group H1(X0,R) of X0 is defined by Ker (∂)
(⊂ C1(X0,R)
).
On the other hand, we define the 0-cochain group and the 1-cochain groupof X0 by
C0(X0,R) := f : V0 −→ R,C1(X0,R) := ω : E0 −→ R |ω(e) = −ω(e),
A CENTRAL LIMIT THEOREM ON CRYSTAL LATTICES 115
respectively. The difference operator d : C0(X0,R) −→ C1(X0,R) is definedby the homomorphism with df(e) := f
(t(e)
)− f
(o(e)
)(e ∈ E0). We also
define H1(X0,R) := C1(X0,R)/Im (d), called the first cohomology group ofX0.
Next we define the transition operator L : C0(X0,R) −→ C0(X0,R) by
Lf(x) := (I − δpd)f(x) =∑
e∈(E0)x
p(e)f(t(e)
) (x ∈ V0, f ∈ C1(X0,R)
),
where the operator δp : C1(X0,R) −→ C0(X0,R) is defined by
δpω(x) := −∑
e∈(E0)x
p(e)ω(e)(x ∈ V0, ω ∈ C1(X0,R)
).
We introduce the quantity γp, called the homological direction of the givenrandom walk on X0, by
γp :=∑e∈E0
m(e)e ∈ C1(X0,R),
where m(e) := p(e)m(o(e)
)(e ∈ E0). It is easy to show that ∂(γp) = 0,
that is, γp ∈ H1(X0,R). It should be noted that the transition probabilityp gives an m-symmetric random walk on X0 if and only if γp = 0. A 1-formω ∈ C1(X0,R) is said to be modified harmonic if
δpω(x) + C1(X0,R)⟨γp, ω⟩C1(X0,R) = 0 (x ∈ V0).
We denote by H1(X0) the space of modified harmonic 1-forms on X0, andequip it with the inner product
⟨⟨ω, η⟩⟩p :=∑e∈E0
m(e)ω(e)η(e)− ⟨γp, ω⟩⟨γp, η⟩(ω, η ∈ H1(X0)
).
Due to the discrete analogue of Hodge–Kodaira theorem (cf. [7, Lemma5.2]), we may identify
(H1(X0), ⟨⟨·, ·⟩⟩p
)with H1(X0,R).
Now let X = (V,E) be a Γ-crystal lattice. Namely, X is a covering graphof a finite graph X0 with an abelian covering transformation group Γ. Wewrite p : E −→ (0, 1] and m : V −→ (0, 1] for the Γ-invariant lifts of p :E0 −→ (0, 1] and m : V0 −→ (0, 1], respectively. Through the covering mapπ : X −→ X0, we take the surjective linear map ρR : H1(X0,R) −→ Γ⊗R(∼=Rd). We consider the transpose tρR : Hom(Γ,R) −→ H1(X0,R), which is ainjective linear map. Here Hom(Γ,R) denotes the space of homomorphismsfrom Γ into R. We induce a flat metric g0 on the Euclidean space Γ ⊗ Rthrough the following diagram:
116 R. NAMBA
(Γ⊗ R, g0) oooo ρR
OO
dual
H1(X0,R)OO
dual
Hom(Γ,R)
tρR
// H1(X0,R) ∼=(H1(X0), ⟨⟨·, ·⟩⟩p
).
This metric g0 is called the Albanese metric on Γ⊗ R.From now on, we realize the crystal lattice X into the continuous model
(Γ⊗ R, g0) in the following manner. A periodic realization of X into Γ⊗ Ris defined by a piecewise linear map Φ : X −→ Γ⊗ R with Φ(σx) = Φ(x) +σ⊗1 (σ ∈ Γ, x ∈ V ). We introduce a special periodic realization Φ0 : X −→Γ⊗ R by
(2.1) Hom(Γ,R)⟨ω,Φ0(x)
⟩Γ⊗R =
∫ x
x∗
ω(x ∈ V, λ ∈ Hom(Γ,R)
),
where x∗ is a fixed reference point satisfying Φ0(x∗) = 0 and ω is the lift ofω to X. Here ∫ x
x∗
ω =
∫cω :=
n∑i=1
ω(e)
for a path c = (e1, . . . , en) with o(e1) = x∗ and t(en) = x. It should be notedthat this line integral does not depend on the choice of a path c. The periodicrealization Φ0 given by above enjoys the so-called modified harmonicity inthe sense that
LΦ0(x)− Φ0(x) = ρR(γp) (x ∈ V ).
We note that this equation is also written as (1.3). Further, such a realizationis uniquely determined up to translation. We call the quantity ρR(γp) theasymptotic direction of the given random walk. We should emphasize thatγp = 0 implies ρR(γp) = 0. However, the converse does not always hold,that is, there is a case γp = 0 and ρR(γp) = 0. (See Subsection 4.2, for anexample.) If we equip Γ ⊗ R with the Albanese metric, then the modifiedharmonic realization Φ0 : X −→ (Γ⊗R, g0) is said to be a modified standardrealization.
3. Proofs of the main results
3.1. A measure–change technique. In what follows, we write λ[x]Γ⊗R :=
Hom(Γ,R)⟨λ,x⟩Γ⊗R (λ ∈ Hom(Γ,R), x ∈ Γ⊗ R) and
dΦ0(e) := Φ0
(t(e)
)− Φ0
(o(e)
)(e ∈ E),
for brevity. Take an orthonormal basis
ω1, . . . , ωd ⊂ Hom(Γ,R)(⊂ (H1(X0), ⟨⟨·, ·⟩⟩p)
)
A CENTRAL LIMIT THEOREM ON CRYSTAL LATTICES 117
and denote by v1, . . . ,vd its dual basis in Γ ⊗ R. Namely, ωi[vj ]Γ⊗R =δij (1 ≤ i, j ≤ d). Then, v1, . . . ,vd is an orthonormal basis in Γ⊗ R withrespect to the Albanese metric g0. We may identify λ = λ1ω1+ · · ·+λdωd ∈Hom(Γ,R) with (λ1, . . . , λd) ∈ Rd. Furthermore, we write xi := ωi[x]Γ⊗R,Φ0(x)i := ωi[Φ0(x)]Γ⊗R and ∂i := ∂/∂λi (i = 1, . . . , d, x ∈ V ). We denoteby O(·) the Landau symbol.
At the beginning, consider the function F = Fx(λ) : V0×Hom(Γ,R) −→ Rdefined by (1.4). We easily see that F = Fx(λ) is a positive function onV0 × Hom(Γ,R) with Fx(0) = 1 (x ∈ V0). In our setting, the followinglemma plays a siginificant role so as to obtain Theorem 1.3.
Lemma 3.1. For every x ∈ V0, the function Fx(·) : Hom(Γ,R) −→ (0,∞)has a unique minimizer λ∗ = λ∗(x).
Proof. Fix a fixed x ∈ V0, we have
∂iFx(λ) = ∂i
( ∑e∈(E0)x
p(e) exp(λ[dΦ0(e)
]Γ⊗R
))
= ∂i
( ∑e∈(E0)x
p(e) exp( d∑
i=1
λi · ωi
[dΦ0(e)
]Γ⊗R
))
=∑
e∈(E0)x
p(e) exp(λ[dΦ0(e)
]Γ⊗R
)dΦ0(e)i(i = 1, . . . , d, λ ∈ Hom(Γ,R)
).
In other words,(∂1Fx(λ), . . . , ∂dFx(λ)
)=
∑e∈(E0)x
p(e) exp(λ[dΦ0(e)
]Γ⊗R
)dΦ0(e) (∈ Γ⊗ R)(3.1)
for λ ∈ Hom(Γ,R). Repeating the above calculation, we have
∂i∂jFx(λ) =∑
e∈(E0)x
p(e) exp(λ[dΦ0(e)
]Γ⊗R
)dΦ0(e)idΦ0(e)j
for λ ∈ Hom(Γ,R) and 1 ≤ i, j ≤ d. Then, we know that(∂i∂jFx(·)
)1≤i,j≤d
,
the Hessian matrix of the function Fx(·), is positive definite. Indeed, con-sider the quadratic form corresponding to the Hessian matrix. Since∑
1≤i,j≤d
∑e∈(E0)x
p(e) exp(λ[dΦ0(e)
]Γ⊗R
)dΦ0(e)idΦ0(e)jξiξj
118 R. NAMBA
=∑
e∈(E0)x
p(e) exp(λ[dΦ0(e)
]Γ⊗R
) d∑i=1
dΦ0(e)iξi
2≥ 0(3.2)
for ξ = (ξ1, . . . , ξd) ∈ Rd and the transition probability p is positive, weeasily see that the Hessian matrix is non-negative definite. By multiplyingboth sides of (3.2) by m(x) and taking the sum which runs over all verticesof X0, it readily follows that
∑e∈E0
m(e) exp(λ[dΦ0(e)
]Γ⊗R
) d∑i=1
dΦ0(e)iξi
2≥ 0
for ξ = (ξ1, . . . , ξd) ∈ Rd. Next suppose that the left-hand side of (3.2) iszero. Then we have
d∑i=1
dΦ0(e)iξi = 0
for all e ∈ E0. This equation implies ⟨Φ0(x), ξ⟩Rd = ⟨Φ0(y), ξ⟩Rd for allx, y ∈ V , where ⟨·, ·⟩Rd stands for the standard inner product on Rd. Letσ1, . . . , σd be generators of Γ ∼= Zd. It follows from the periodicity of Φ0
that ⟨σi, ξ⟩Rd = 0 (1 ≤ i ≤ d). Hence we conclude ξ = 0. Namely, we haveproved the positive definiteness of the Hessian matrix.
This implies that the function Fx(·) : Hom(Γ,R) −→ (0,∞) is strictlyconvex for every x ∈ V0. Moreover, it is easily observed that
lim|λ|Rd→∞
Fx(λ) = ∞ (x ∈ X0),
due to its definition. Consequently, we know that there exists a uniqueminimizer λ∗ = λ∗(x) ∈ Hom(Γ,R) of Fx(λ) for each x ∈ V0, therebycompleting the proof.
Now consider the positive function p : E0 −→ (0, 1] given by (1.5). Bydefinition, we easily see that the function p also gives a positive transitionprobability on X0. Thus the transition probability p : E0 −→ (0, 1] yields an
irreducible random walk (Ωx∗(X), Px∗ , w(p)n ∞n=0) with values in X. Apply-
ing the Perron-Frobenius theorem again, we find a unique positive functionm : V0 −→ (0, 1] satisfying (1.2). Put m(e) := p(e)m(o(e)) (e ∈ E0). Wealso denote by p : E −→ (0, 1] and m : V −→ (0, 1] the Γ-invariant liftsof p : E0 −→ (0, 1] and m : V0 −→ (0, 1], respectively. As in the previous
section, we construct the (p-)Albanese metric g(p)0 on Γ⊗R associated with
the transition probability p. We take an orthonormal basis ω(p)1 , . . . , ω
(p)d
in Hom(Γ,R)(⊂ (H1(X0), ⟨⟨·, ·⟩⟩p)
).
A CENTRAL LIMIT THEOREM ON CRYSTAL LATTICES 119
We introduce the transition operator L(p) : C∞(X) −→ C∞(X) associatedwith the transition probability p by
L(p)f(x) :=∑e∈Ex
p(e)f(t(e)) (x ∈ V ).
Recalling (3.1) and the definition of λ∗ = λ∗(x), we see that(∂1Fx
(λ∗(x)
), . . . , ∂dFx
(λ∗(x)
))=
∑e∈(E0)x
p(e) exp(λ∗(x)
[dΦ0(e)
]Γ⊗R
)dΦ0(e) = 0
holds for every x ∈ V0. This immediately leads to
(3.3) L(p)Φ0(x)− Φ0(x) =∑e∈Ex
p(e)dΦ0(e) = 0 (x ∈ V ).
From this equation, one concludes that the given p-modified standard real-ization Φ0 : X −→ (Γ⊗R, g0) in the sense of (1.3) is the harmonic realizationassociated with the changed transition probability p.
Remark. Equation (3.3) implies ρR(γp) = 0. We also emphasize that thetransition probability p : E0 −→ (0, 1] coincides with the original one p :E0 −→ (0, 1] provided that ρR(γp) = 0.
Remark. In our setting, it is essential to assume that the given transitionprobability p is positive. Because, if it were not for the positivity of p, theassertion of Lemma 3.1 would not hold in general. (There is a case where thefunction Fx(·) has no minimizers.) On the other hand, to obtain Theorems1.1 and 1.2, it is sufficient to impose that the given transition probability pis non-negative with p(e) + p(e) > 0 (e ∈ E).
3.2. Proofs of Theorems 1.3 and 1.4. This subsection is devoted toproofs of Theorems 1.3 and 1.4. Following the argument as in [4, Theorem2.2] for the random walk associated with the changed transition probabilityp, we can carry out the proof of Theorem 1.3. Though a minor change ofthe proof is required, the argument is a little bit easier due to ρR(γp) = 0.
As the first step, we prove the following lemma.
Lemma 3.2. For any f ∈ C∞0 (Γ ⊗ R), as N → ∞, ε 0 and N2ε 0,
we have ∥∥∥ 1
Nε2(I − LN
(p))Pεf − Pε
(∆(p)
2f)∥∥∥X
∞−→ 0.
Here Pε : C∞(Γ ⊗ R) −→ C∞(X) (0 ≤ ε ≤ 1) is a scaling operator definedby
Pεf(x) := f(εΦ0(x)
)(x ∈ X)
120 R. NAMBA
and ∆(p) stands for the positive Laplacian −∑d
i=1(∂2/∂x2i ) on Γ⊗ R asso-
ciated with the p-Albanese metric g(p)0 .
Proof. We define the function AN (Φ0)ij : V −→ R (i, j = 1, . . . , d, N ∈ N)by
AN (Φ0)ij(x)
:=∑
c∈Ωx,N (X)
p(c)(Φ0
(t(c))− Φ0(x)
)i
(Φ0
(t(c))− Φ0(x)
)j
(x ∈ V ),
where p(c) := p(e1) · · · p(eN ) for c = (e1, . . . , eN ) ∈ Ωx,N (X). ApplyingTaylor’s expansion formula, we have
(I − LN(p))Pεf(x) = −ε
d∑i=1
∂f
∂xi
(εΦ0(x)
) ∑c∈Ωx,N (X)
p(c)(Φ0
(t(c))− Φ0(x)
)i
− ε2
2
∑1≤i,j≤d
∂2f
∂xi∂xj
(εΦ0(x)
)AN (Φ0)ij(x) +O
((Nε)3
).(3.4)
We see that the first term of the right-hand side of (3.4) equals 0 due to thep-harmonicity of Φ0. Next we define the function A(Φ0)ij : V0 −→ R (i, j =1, . . . , d) by
A(Φ0)ij(x) :=∑
e∈(E0)x
p(e)dΦ0(e)idΦ0(e)j (x ∈ V ).
We note that A(Φ0)ij(π(x)
)= A1(Φ0)ij(x) (x ∈ V, i, j = 1, . . . , d) because
AN (Φ0)ij is Γ-invariant. Then, using the p-harmonicity again,
AN (Φ0)ij(x) =N−1∑k=0
Lk(p)
(A(Φ0)ij
)(π(x)
)(x ∈ V ).
Applying the ergodic theorem for L(p) (cf. [4, Theorem 3.2]), we have
1
N
N−1∑k=0
Lk(p)
(A(Φ0)ij
)(π(x)
)=∑x∈V0
m(x)A(Φ0)ij(x) +O( 1
N
).
Then, (2.1) and p-harmonicity of Φ0 imply∑x∈V0
m(x)A(Φ0)ij(x) =∑e∈E0
m(e)ω(p)i (e)ω
(p)j (e) = ⟨⟨ω(p)
i , ω(p)j ⟩⟩p = δij
for 1 ≤ i, j ≤ d. Putting it all together, we obtain
1
Nε2(I − LN
(p))Pεf = Pε
(∆(p)
2f)+O(N2ε) +O
( 1
N
).
A CENTRAL LIMIT THEOREM ON CRYSTAL LATTICES 121
Finally, letting N → ∞, ε 0 and N2ε 0, we complete the proof. Lemma 3.2 immediately leads to the following lemma. (See [4, Theorem
2.1 and Lemma 4.2] for details.)
Lemma 3.3. (1) For any f ∈ C∞(Γ⊗ R), and 0 ≤ s ≤ t, we have
limn→∞
∥∥∥L[nt]−[ns](p) Pn−1/2f − Pn−1/2e−(t−s)∆(p)/2f
∥∥∥X∞
= 0.
(2) We fix 0 ≤ t1 < · · · < tℓ < ∞ (ℓ ∈ N). Then,
(X(n)t1
, . . . ,X(n)tℓ
)(d)−→ (B
(p)t1
, . . . , B(p)tℓ
) (n → ∞),
where (B(p)t )t≥0 is a (Γ ⊗ R, g(p)0 )-valued standard Brownian motion with
B(p)0 = 0.
Having obtained Lemma 3.3, it is sufficient to show the tightness ofP(n)∞n=1 for completing the proof of Theorem 1.3.
Lemma 3.4. The sequence P(n)∞n=1 is tight in C0
([0,∞), (Γ⊗ R, g(p)0 )
).
Proof. Throughout the proof, C denotes a positive constant that maychange at every occurrence. We put ∥dΦ0∥∞ := maxe∈E0 ∥dΦ0(e)∥g(p)0
.
By virtue of the celebrated Kolmogorov’s criterion, it is sufficient to showthat there exists some positive constant C independent of n such that
(3.5) EPx∗[∥∥X(n)
t − X(n)s
∥∥4g(p)0
]≤ C(t− s)2 (0 ≤ s ≤ t, n ∈ N).
We split the proof into two cases: (I) : t− s < n−1, (II) : t− s ≥ n−1.First we consider the case (I). In both cases ns ≥ [nt] and ns < [nt], we
have∥∥X(n)t − X(n)
s
∥∥g(p)0
≤ n1/2(t− s)∥∥ξ(p)[nt]+1 − ξ
(p)[nt]∥g(p)0
+∥∥ξ(p)[nt] − ξ
(p)[nt]−1∥g(p)0
≤ 2∥dΦ0∥∞n1/2(t− s).
Noting n2(t− s)2 < 1, we obatin the desired estimate (3.5) for case (I).Next we consider the case (II). Let F be the fundamental domain in X
containing x∗ ∈ V and define Mℓi = Mℓ
i(Φ0) : V −→ R (i = 1, . . . , d, ℓ =1, 2, 3, 4) by
Mℓi(x) :=
∑e∈Ex
p(e)dΦ0(e)ℓi (x ∈ V ).
We note that Mℓi is Γ-invariant and ∥Mℓ
i∥X∞ ≤ ∥dΦ0∥ℓ∞ (i = 1, . . . , d). More-over, we obtain M1
i ≡ 0 (i = 1, . . . , d) due to the p-harmonicity of Φ0.
122 R. NAMBA
Here we give a bound on EPx∗[∥∥X (n)
Mn
− X (n)Nn
∥∥4g(p)0
](n ∈ N, M ≥ N ∈ N).
First of all, we have
EPx∗[∥∥X (n)
Mn
−X (n)Nn
∥∥4g(p)0
]≤ Cn−2 max
i=1,...,dmaxx∈F
∑c∈Ωx,M−N (X)
p(c)(Φ0
(t(c))− Φ0(x)
)4i
.(3.6)
Now fix i = 1, . . . , d and x ∈ F . For k = 1, . . . ,M −N , we have∑c∈Ωx,k(X)
p(c)(Φ0
(t(c))− Φ0(x)
)4i
=∑
c′∈Ωx,k−1(X)
p(c′)∑
e∈Et(c′)
p(e)
×(
Φ0
(t(e)
)− Φ0
(o(e)
))i+(Φ0
(o(e)
)− Φ0(x)
)i
=
∑c′∈Ωx,k−1(X)
p(c′)M4i
(t(c′)
)+ 4
∑c′∈Ωx,k−1(X)
p(c′)(Φ0
(t(c′)
)− Φ0(x)
)iM3
i
(t(c′)
)+ 6
∑c′∈Ωx,k−1(X)
p(c′)(Φ0
(t(c′)
)− Φ0(x)
)2iM2
i
(t(c′)
)+
∑c′∈Ωx,k−1(X)
p(c′)(Φ0
(t(c′)
)− Φ0(x)
)4i
≤∑
c′∈Ωx,k−1(X)
p(c′)(Φ0
(t(c′)
)− Φ0(x)
)4i+ ∥M4
i ∥X∞
+ 4∥dΦ0∥∞∥M3i ∥X∞ + 6∥M2
i ∥X∞∑
c∈Ωx,k−1(X)
p(c)(Φ0
(t(c))− Φ0(x)
)2i.(3.7)
Moreover, the p-harmonicity implies∑c∈Ωx,k−1(X)
p(c)(Φ0
(t(c))− Φ0(x)
)2i
=∑
c′∈Ωx,k−2(X)
p(c′)(
Φ0
(t(c′)
)− Φ0(x)
)2i+M2
i
(t(c′)
)≤ (k − 1)∥dΦ0∥2∞.(3.8)
A CENTRAL LIMIT THEOREM ON CRYSTAL LATTICES 123
It follows from (3.7) and (3.8) that
∑c∈Ωx,k(X)
p(c)(Φ0
(t(c))− Φ0(x)
)4i
≤∑
c∈Ωx,k−1(X)
p(c)(Φ0
(t(c))− Φ0(x)
)4i
+ 5∥dΦ0∥4∞ + 6∥dΦ0∥4∞(k − 1)
≤ Ck2.(3.9)
Putting k = M −N , M = [nt] + i, N = [ns] + j (i, j = 0, 1) and combining(3.6) with (3.9), we obtain
EPx∗[∥∥X(n)
t − X(n)s
∥∥4g(p)0
]≤ EPx∗
[maxi,j=0,1
∥∥X (n)[nt]+i
n
−X (n)[ns]+j
n
∥∥4g(p)0
]≤ Cn−2
([nt]− [ns] + 1
)2≤ C
(t− s+
2
n
)2≤ C
(t− s) + 2(t− s)
2≤ C(t− s)2,
where we used [nt]− [ns] ≤ n(t− s)+1 and n−1 ≤ t− s. Therefore, we haveshown the the desired estimate (3.5) for case (II).
Next we prove Theorem 1.4.
Proof of Theorem 1.4. For n ∈ N and x, y ∈ V , we have
p(n,x, y) =∑
(e1,...,en)∈Ωx,n(X)o(e1)=x, t(en)=y
p(e1) · · · p(en)
=∑
(e1,...,en)∈Ωx,n(X)o(e1)=x, t(en)=y
p(e1) · · · p(en) · exp( n∑
i=1
λ∗(o(ei)
)[dΦ0(ei)]Γ⊗R
)
× Fo(e1)
(λ∗(o(e1))
)−1 · · ·Fo(en)
(λ∗(o(en))
)−1
=∑
(e1,...,en)∈Ωx,n(X)o(e1)=x, t(en)=y
p(e1) · · · p(en)
× exp( n∑
i=1
λ∗(o(ei)
)[dΦ0(ei)]Γ⊗R −
n∑i=1
logFo(ei)
(λ∗(o(ei))
)).
124 R. NAMBA
Using the ergodic theorem for 1-chains (cf. [7]):
(3.10)1
n
n∑i=1
f(ei) =∑e∈E0
m(e)f(e) +O( 1n
)(f : E0 −→ R),
we obtain
1
n
n∑i=1
(λ∗(o(ei)
)[dΦ0(ei)]Γ⊗R − logFo(ei)
(λ∗(o(ei))
))=∑e∈E0
m(e)(λ∗(o(e)
)[dΦ0(e)]Γ⊗R − logFo(e)
(λ∗(o(e))
))+O
( 1n
)=∑x∈V0
m(x)(λ∗(x)
[ ∑e∈(E0)x
dΦ0(e)]Γ⊗R
− logFx
(λ∗(x)
))+O
( 1n
)=∑x∈V0
m(x)(λ∗(x)
[ρR(γp)
]Γ⊗R − logFx
(λ∗(x)
))+O
( 1n
)for x, y ∈ V . Here we used the p-modified harmonicity of Φ0 for the finalline. Finally, we obtain
p(n, x, y) = p(n, x, y) exp(n∑x∈V0
m(x)
×(λ∗(x)
[ρR(γp)
]Γ⊗R − logFx
(λ∗(x)
))+O(1)
)for x, y ∈ V . This completes the proof.
Remark. Let us consider a special case where the Γ-crystal lattice X is givenby a covering graph of an ℓ-bouquet graph (ℓ ∈ N) consisting of one vertexx ∈ V0 and ℓ-loops. Without using the ergodic theorem (3.10) in the proofof Theorem 1.4, we also obtain
p(n, x, y) =∑
(e1,...,en)∈Ωx,n(X)o(e1)=x, t(en)=y
p(e1) · · · p(en)
× exp( n∑
i=1
λ∗(x)[dΦ0(ei)]Γ⊗R
)· Fx
(λ∗(x)
)−n
= p(n, x, y) exp(λ∗(x)
[Φ0(y)− Φ0(x)
]Γ⊗R
)· Fx
(λ∗(x)
)−n(3.11)
for every n ∈ N, x, y ∈ V .
A CENTRAL LIMIT THEOREM ON CRYSTAL LATTICES 125
3.3. A relationship to a discrete analogue of Girsanov’s theorem.In this subsection, we discuss a relationship between our formula (3.11) anda discrete analogue of Girsanov’s theorem due to Fujita [3].
Let X = (V,E) be a crystal lattice covered with a one-bouquet graphX0 = (V0, E0); V0 = x and E0 = e, e, by the group action Γ = ⟨σ⟩ ∼= Z1.We consider a random walk on X0 with the transition probability
p(e) = p and p(e) = 1− p (0 < p < 1).
We introduce a bijective linear map ρR : H1(X0,R) −→ Γ ⊗ R(∼= R1) byρR([e]) = σ. Then we have γp = (2p − 1)[e] and ρR(γp) = (2p − 1)σ. Letu ⊂ Hom(Γ,R) =
(H1(X0,R), ⟨⟨·, ·⟩⟩p
)be a dual basis of σ ⊗ 1 = σ ⊂
Γ⊗R. We easily see that ⟨⟨u, u⟩⟩p = 4p(1− p). Hence the orthogonalizationv ⊂ Hom(Γ,R) of u is given by
v =1√
4p(1− p)u.
To the end, we identify λv ∈ Hom(Γ,R) with λ ∈ R. We denote by v ⊂Γ⊗R the dual basis of v. Then we observe that the realization Φ0 : X −→(Γ⊗ R; v) defined by
dΦ0(e) := σ =1√
4p(1− p)v
is the modified standard realization of X.We now consider the function F = Fx(λ) defined by (1.4), that is,
Fx(λ) = p exp( λ√
4p(1− p)
)+ (1− p) exp
(− λ√
4p(1− p)
)(λ ∈ R).
Then we know that the minimizer λ∗ = λ∗(x) and Fx(λ∗) are given by
λ∗ =√p(1− p) log
p− 1
p, Fx(λ∗) =
√4p(1− p).
We fix x ∈ V satisfying Φ0(x) = 0. For y ∈ V , we write Φ0(y) = k(y)v.Then the formula (3.11) implies
p(n, x, y) = p(n, x, y) ·(p− 1
p
)−k(y)/2·(√
4p(1− p))−n
(n ∈ N, y ∈ V ).
In [3, page 115], the above formula is called a discrete analogue of Girsanov’stheorem for a non-symmetric random walk Zn∞n=0 on Z1 given by the sumof independent random variables ξi∞i=1 with P(ξi = 1) = p and P(ξi =−1) = 1 − p (i = 1, 2, . . . ). Hence we may regard the formula (3.11) as ageneralization of the above discrete Girsanov’s theorem to the case of non-symmetric random walks on the ℓ-bouquet graph.
126 R. NAMBA
4. Examples
In [4, Section 7], several examples of the modified standard realizationof crystal lattices associated with the non-symmetric random walks are dis-cussed. In this final section, we give two concrete examples of non-symmetricrandom walks on crystal lattices and calculate the changed transition prob-ability p for each example.
4.1. The hexagonal lattice. In this subsection, we consider the hexago-nal lattice, as a typical example of crystal lattices. Let X = (V,E) be ahexagonal lattice, V = Z2 = x = (x1, x2) |x1, x2 ∈ Z and
E =(x,y) ∈ V 2
∣∣x− y = ±(1, 0), x− y = (0, (−1)x1+x2).
(See Figure 1). We introduce a non-symmetric random walk on X in thefollowing way. If x = (x1, x2) ∈ V is a vertex so that x1 + x2 is even, set
p(x,x+ (1, 0)
)=
1
2, p
(x,x− (0, 1)
)=
1
3, p
(x,x− (1, 0)
)=
1
6.
If x1 + x2 is odd, set
p(x,x− (1, 0)
)=
1
6, p
(x,x+ (0, 1)
)=
1
3, p
(x,x+ (1, 0)
)=
1
2.
We see that X is invariant under the action Γ = ⟨σ1, σ2⟩(∼= Z2) generatedby
σ1(x) = x+ (1, 1), σ2(x) = x+ (−1, 1) (x ∈ V ).
X = (V,E)
!e2
!e1!e3 !x1
!x2
σ1σ2
π
Γ = ⟨σ1,σ2⟩
e1e2e3
x2
x1
X0 = (V0, E0)
1
Figure 1. Hexagonal lattice and its quotient.
The quotient graph X0 = Γ\X is a finite graph X0 = (V0, E0) consistingof two vertices x1,x2 with three multiple edges E0 = (E0)x1 ∪ (E0)x2 =e1, e2, e3 ∪ e1, e2, e3.
A CENTRAL LIMIT THEOREM ON CRYSTAL LATTICES 127
We put [c1] := [e1 ∗ e2] and [c2] := [e3 ∗ e2]. Then, the first homologygroup H1(X0,R) is spanned by [c1], [c2]. Solving (1.2), we have m(x1) =m(x2) = 1/2. We define the surjective linear map ρR : H1(X0,R) −→ Γ ⊗R(∼= R2) by ρR([c1]) := σ1, ρR([c2]) := σ2. Thus, the homological directionγp and the asymptotic direction ρR(γp) are given by
γp =1
6[c1]−
1
6[c2], ρR(γp) =
1
6σ1 −
1
6σ2( = 0),
respectively. We will determine the modified standard realization Φ0 :X −→ (Γ ⊗ R, g0). We set x1 = (0, 0), x2 = (0, 1) in V . Without lossof generality, we may put Φ0(x1) = 0 ∈ Γ⊗ R. By (1.3), we have
Φ0(x1) = 0, Φ0(x2) = −1
3σ1 −
1
3σ2.
Now let v1, v2 be an orthonormal basis in Hom(Γ,R)(⊂ (H1(X0), ⟨⟨·, ·⟩⟩p)
)and v1,v2 its dual basis in (Γ⊗ R, g0). Then, we have
σ1 =6√7
7v1 −
9√70
70v2, σ2 =
3√70
10v2
with the Albanese metric
⟨σ1, σ1⟩g0 =63
10, ⟨σ1, σ2⟩g0 =
27
10, ⟨σ2, σ2⟩g0 =
63
10,
by following the computations in [4, Subsection 7.3]. Hence, we find that themodified standard realization Φ0 : X −→ (Γ⊗R, g0) is given by Φ0(x1) = 0and
Φ0(x2) = −2√7
7v1 −
√70
7v2.
Now we are in a position to consider the function F introduced in (1.4).We identify λ ∈ Hom(Γ,R) with (λ1, λ2) ∈ R2. Then, we have
Fx1(λ)
=1
2exp
(4√7
7λ1 −
19√70
70λ2
)+
1
3exp
(− 2
√7
7λ1 −
√70
7λ2
)+
1
6exp
(− 2
√7
7λ1 +
11√70
70λ2
),
Fx2(λ)
=1
6exp
(− 4
√7
7λ1 +
19√70
70λ2
)+
1
3exp
(2√7
7λ1 +
√70
7λ2
)+
1
2exp
(2√7
7λ1 −
11√70
70λ2
).
(4.1)
128 R. NAMBA
Differentiating both sides of (4.1) with respect to λ1 and λ2, we have
∂1Fx1(λ) =2√7
7exp
(4√7
7λ1 −
19√70
70λ2
)− 2
√7
21exp
(− 2
√7
7λ1 −
√70
7λ2
)−
√7
21exp
(− 2
√7
7λ1 +
11√70
70λ2
),
∂2Fx1(λ) = −19√70
140exp
(4√7
7λ1 −
19√70
70λ2
)−
√70
21exp
(− 2
√7
7λ1 −
√70
7λ2
)+
11√70
420exp
(− 2
√7
7λ1 +
11√70
70λ2
),
∂1Fx2(λ) = −2√7
21exp
(− 4
√7
7λ1 +
19√70
70λ2
)+
2√7
21exp
(2√7
7λ1 +
√70
7λ2
)+
√7
7exp
(2√7
7λ1 −
11√70
70λ2
),
∂2Fx2(λ) =19√70
420exp
(− 4
√7
7λ1 +
19√70
70λ2
)+
√70
21exp
(2√7
7λ1 +
√70
7λ2
)− 11
√70
140exp
(2√7
7λ1 −
11√70
70λ2
).
To find the minimizers of functions Fx1(·) and Fx2(·), it is sufficient to solvethe following two algebraic equations:
∂1Fx1(λ1, λ2) = 0
∂2Fx1(λ1, λ2) = 0,
∂1Fx2(λ1, λ2) = 0
∂2Fx2(λ1, λ2) = 0.
Solving these equations, the minimizers λ∗(x1), λ∗(x2) of Fx1(·), Fx2(·) aregiven by
λ∗(x1) =(√7
14log 26 +
√7
6log
14
3,
√70
21log 26
),
λ∗(x2) =(√7
14log
3
26−
√7
6log 14,
√70
21log
3
26
),
(4.2)
A CENTRAL LIMIT THEOREM ON CRYSTAL LATTICES 129
respectively. Then, it follows from (4.1) and (4.2) that
Fx1
(λ∗(x1)
)= 7 · 26−13/21 ·
(143
)−1/3,
Fx2
(λ∗(x2)
)=
21
26· 14−1/3 ·
( 3
26
)−8/21.
Finally, we determine the changed transition probability p by
p(e1) =1
3, p(e2) =
1
21, p(e3) =
13
21,
p(e1) =1
3, p(e2) =
1
21, p(e3) =
13
21.
Then, the invariant measure m : x1,x2 −→ (0, 1] is also given by m(x1) =m(x2) = 1/2. Hence we know that the random walk associated with thechanged transition probability p is m-symmetric, that is,
p(e)m(o(e)
)= p(e)m
(t(e)
)(e ∈ E0).
It automatically implies γp = 0 and also ρR(γp) = 0.
4.2. The dice lattice. In this subsection, we discuss a non-symmetric ran-dom walk on an infinite graph called the dice lattice or the dice graph. Thedice graph X = (V,E) is one of abelian covering graphs which has a free ac-tion by the lattice group Γ ∼= Z2 generated by σ1, σ2, and the correspondingquotient graph X0 = (V0, E0) := Γ\X is a finite graph consisting of threevertices V0 = x,y, z. (See Figure 2 or the description in [9].) In view ofthe shape of the quotient graph X0, we may regard the dice lattice X assomething like a “hybrid ” of a triangular lattice and a hexagonal lattice.
From now on, we consider a non-symmetric random walk on X by givingthe transition probability on the quotient X0 in the following way. We set
p(e1) =1
4, p(e2) =
1
6, p(e3) =
1
12, p(e4) =
1
4, p(e5) =
1
6, p(e6) =
1
12,
p(e1) =1
6, p(e2) =
1
3, p(e3) =
1
2, p(e4) =
1
6, p(e5) =
1
3, p(e6) =
1
2.
Solving (1.2), we have m(x) = 1/2, m(y) = m(z) = 1/4.Next we define four 1-cycles [c1], [c2], [c3], [c4] on X0 by
[c1] := [e1 ∗ e2], [c2] := [e3 ∗ e2], [c3] := [e4 ∗ e5], [c4] := [e6 ∗ e5].
Then [c1], [c2], [c3], [c4] spans the first homology group H1(X0,R). Wedefine the linear map ρR from H1(X0,R) onto Γ⊗ R ∼= R2 by
ρR([c1]) = σ1 − σ2, ρR([c2]) = −σ2, ρR([c3]) = σ2, ρR([c4]) = σ2 − σ1.
130 R. NAMBA
Then, the homological direction γp and the asymptotic direction ρR(γp) ofthe random walk on X0 are calculated as
γp =1
12
([c1]− [c2] + [c3]− [c4]
),
ρR(γp) =1
12
(σ1 − σ2)− (−σ2) + σ2 − (σ2 − σ1)
=
1
6σ1( = 0),
respectively.
!e1
!e2
!e3
!e4!e6
!e5
σ1
σ2
z
y
e1e2e3
e4e5e6
x
X = (V,E)
X0 = (V0, E0)
π
Γ = ⟨σ1,σ2⟩
1
Figure 2. Dice lattice and its quotient.
Now we determine the modified standard realization Φ0 : X −→ Γ ⊗ R.Here we may put Φ0
(o(ei)
)= 0 (i = 1, 2, 3, 4, 5, 6), without loss of generality.
Noting (1.3) and the group action Γ, we have
Φ0
(t(e1)
)=
2
3σ1 −
1
3σ2, Φ0
(t(e2)
)= −1
3σ1 +
2
3σ2,
Φ0
(t(e3)
)= −1
3σ1 −
1
3σ2, Φ0
(t(e4)
)=
1
3σ1 +
1
3σ2,(4.3)
Φ0
(t(e5)
)=
1
3σ1 −
2
3σ2, Φ0
(t(e6)
)= −2
3σ1 +
1
3σ2.
A CENTRAL LIMIT THEOREM ON CRYSTAL LATTICES 131
Let ω1, ω2, ω3, ω4 ⊂ H1(X0,R) be the dual basis of [c1], [c2], [c3], [c4],that is, ωi([cj ]) = δij (1 ≤ i, j ≤ 4). Recalling that each ωi is a modifiedharmonic 1-form, we have
ω1(e1) =3
4, ω1(e2) = ω1(e3) = −1
4, ω1(e4) = ω1(e5) = ω1(e6) = − 1
12,
ω2(e1) = ω2(e2) = − 5
12, ω2(e3) =
7
12, ω2(e4) = ω2(e5) = ω2(e6) =
1
12,
ω3(e1) = ω3(e2) = ω3(e3) = − 1
12, ω3(e4) =
3
4, ω3(e5) = ω3(e6) = −1
4,
ω4(e1) = ω4(e2) = ω4(e3) =1
12, ω4(e4) = ω4(e5) = − 5
12, ω4(e6) =
7
12.
Then, the direct computation gives us
⟨⟨ω1, ω1⟩⟩ =1
9, ⟨⟨ω1, ω2⟩⟩ = − 1
18,
⟨⟨ω1, ω3⟩⟩ = − 1
72, ⟨⟨ω1, ω4⟩⟩ =
1
72,
⟨⟨ω2, ω2⟩⟩ =1
9, ⟨⟨ω2, ω3⟩⟩ =
1
72,
⟨⟨ω2, ω4⟩⟩ = − 1
72, ⟨⟨ω3, ω3⟩⟩ =
1
9,
⟨⟨ω3, ω4⟩⟩ = − 1
18, ⟨⟨ω4, ω4⟩⟩ =
1
9.
We introduce the basis u1, u2 in Hom(Γ,R) by the dual of σ1⊗1, σ2⊗1 ⊂Γ⊗ R. Since the dice graph X is a non-maximal abelian covering graph ofX0 with Γ ∼= ⟨σ1, σ2⟩, we need to find a Z-basis of the lattice
L =ω ∈ H1(X0,R)
∣∣ω([c]) = 0 for every cycle c on X.
It is easy to find that
u1 =tρR(u1) = ω1 − ω4, u2 =
tρR(u2) = −ω1 − ω2 + ω3 + ω4
form a Z-basis of the lattice L. Carrying out the direct computation again,we have
(4.4) ⟨⟨u1, u1⟩⟩ =7
36, ⟨⟨u1, u2⟩⟩ = −1
9, ⟨⟨u2, u2⟩⟩ =
2
9.
Let v1, v2 be the Gram–Schmidt orthogonalization of the basis u1, u2 ⊂Hom(Γ,R). By (4.4), we have
v1 =6√7
7u1, v2 =
6√70
35u1 +
3√70
10u2.
132 R. NAMBA
We denote by v1,v2 the dual basis of v1, v2 in Γ⊗ R. Then we obtain
(4.5) σ1 =6√7
7v1, σ2 =
6√70
35v1 +
3√70
10v2
Combining (4.5) with (4.3), we finally determine the modified standard re-alization Φ0 : X −→ (Γ⊗ R, g0) ∼= (R2; v1,v2) of X by
Φ0
(t(e1)
)=
20√7− 2
√70
35v1 −
√70
10v2,
Φ0
(t(e2)
)=
−10√7 + 4
√70
35v1 +
√70
5v2,
Φ0
(t(e3)
)=
−10√7− 2
√70
35v1 −
√70
10v2,
Φ0
(t(e4)
)=
10√7 + 2
√70
35v1 +
√70
10v2,
Φ0
(t(e5)
)=
10√7− 4
√70
35v1 −
√70
5v2,
Φ0
(t(e6)
)=
−20√7 + 2
√70
35v1 +
√70
10v2.
(See Figure 3 below.)Now we are in a position to consider the function F : V0×Hom(Γ,R) −→
(0,∞) defined by (1.4). In what follows, we identify λ = λ1v1 + λ2v2 ∈Hom(Γ,R) with (λ1, λ2) ∈ R2. In addition, it follows from Φ0
(o(ei)
)=
0 (i = 1, 2, 3, 4, 5, 6) that
dΦ0(e1) =20√7− 2
√70
35v1 −
√70
10v2,
dΦ0(e2) =−10
√7 + 4
√70
35v1 +
√70
5v2,
dΦ0(e3) =−10
√7− 2
√70
35v1 −
√70
10v2,
dΦ0(e4) =10√7 + 2
√70
35v1 +
√70
10v2,(4.6)
dΦ0(e5) =10√7− 4
√70
35v1 −
√70
5v2,
dΦ0(e6) =−20
√7 + 2
√70
35v1 +
√70
10v2
A CENTRAL LIMIT THEOREM ON CRYSTAL LATTICES 133
Then, by (4.6), we have
Fx(λ1, λ2) =1
4exp
(20√7− 2√70
35λ1 −
√70
10λ2
)+
1
6exp
(−10√7 + 4
√70
35λ1 +
√70
5λ2
)+
1
12exp
(−10√7− 2
√70
35λ1 −
√70
10λ2
)+
1
4exp
(10√7 + 2√70
35λ1 +
√70
10λ2
)+
1
6exp
(10√7− 4√70
35λ1 −
√70
5λ2
)+
1
12exp
(−20√7 + 2
√70
35λ1 +
√70
10λ2
),
Fy(λ1, λ2) =1
6exp
(−20√7 + 2
√70
35λ1 +
√70
10λ2
)+
1
3exp
(10√7− 4√70
35λ1 −
√70
5λ2
)+
1
2exp
(10√7 + 2√70
35λ1 +
√70
10λ2
),
Fz(λ1, λ2) =1
6exp
(−10√7− 2
√70
35λ1 −
√70
10λ2
)+
1
3exp
(−10√7 + 4
√70
35λ1 +
√70
5λ2
)+
1
2exp
(20√7− 2√70
35λ1 −
√70
10λ2
).
By solving the following equations:∂1Fx(λ1, λ2) = 0
∂2Fx(λ1, λ2) = 0,
∂1Fy(λ1, λ2) = 0
∂2Fy(λ1, λ2) = 0,
∂1Fz(λ1, λ2) = 0
∂2Fz(λ1, λ2) = 0,
we find that the minimizers λ∗(x), λ∗(y) and λ∗(z) of functions Fx(·), Fy(·)and Fz(·) are given by
λ∗(x) =(−
√7
6log 3,
4√7−
√70
42log 3
),
λ∗(y) =(−
√7
6log 3,
√70
21log 2 +
2√7−
√70
21log 3
),
λ∗(z) =(−
√7
6log 3,−
√70
21log 2 +
2√7
21log 3
),
134 R. NAMBA
respectively, and
Fx
(λ∗(x)
)=
√3 + 1
3, Fy
(λ∗(y)
)= Fz
(λ∗(z)
)= 3 · 6−2/3.
v2
v1σ1
σ2
Φ0!t("e1)
#
Φ0!t("e2)
#
Φ0!t("e3)
#
Φ0!t("e4)
#
Φ0!t("e5)
#
Φ0!t("e6)
#
ρR(γp)
1
Figure 3. Modified standard realization of the dice lattice.
Finally, we determine the changed transition probability p by
p(e1) =3−
√3
8, p(e2) =
√3− 1
4, p(e3) =
3−√3
8,
p(e4) =3−
√3
8, p(e5) =
√3− 1
4, p(e6) =
3−√3
8,
p(e1) = p(e2) = p(e3) = p(e4) = p(e5) = p(e6) =1
3.
Then, the invariant measure m is given by m(x) = 1/2 and m(y) = m(z) =1/4. Moreover, the homological direction γp and the asymptotic directionρR(γp) of the random walk associated with p are given by
γp =5− 3
√3
48
([c1] + [c2] + [c3] + [c4]
)(= 0), ρR(γp) = 0,
A CENTRAL LIMIT THEOREM ON CRYSTAL LATTICES 135
respectively. Then it follows from Theorem 1.4 that there exist some positiveconstants C1 and C2 such that
C1p(n, x, y)(
6√12(
√3− 1)
)n ≤ p(n, x, y) ≤ C2p(n, x, y)(
6√12(
√3− 1)
)nfor all n ∈ N and x, y ∈ V .
Acknowledgement. The author would like to thank his advisor Profes-sor Hiroshi Kawabi for useful discussion and constant encouragement. Heexpresses his gratitude to Professors Satoshi Ishiwata, Tomoyuki Kakehi,Atsushi Katsuda and Ryokichi Tanaka for valuable advice and comments.He also would like to thank the anonymous referee for providing helpfulcomments and suggestions.
References
[1] G. Alexopoulos: Random walks on discrete groups of polynomial growth, Ann. Probab.30 (2002), pp. 723–801.
[2] R. Durrett: Probability: Theory and Examples, Cambridge Series in Statistical andProbabilistic Mathematics, Fourth Edition, Cambridge University Press, 2010.
[3] T. Fujita: Random Walks and Stochastic Analysis – From Gambling to MathematicalFinance (in Japanese), Nihon Hyoronsya, 2008.
[4] S. Ishiwata, H. Kawabi and M. Kotani: Long time asymptotics of non-symmetricrandom walks on crystal lattices, J. Funct. Anal. 272 (2017), pp.1553–1624.
[5] M. Kotani: A central limit theorem for magnetic transition operators on a crystallattice, J. London Math. Soc. 65 (2002), pp. 464–482.
[6] M. Kotani and T. Sunada: Albanese maps and off diagonal long time asymptotics forthe heat kernel, Comm. Math. Phys. 209 (2000), pp. 633–670.
[7] M. Kotani and T. Sunada: Large deviation and the tangent cone at infinity of acrystal lattice, Math. Z. 254 (2006), pp. 837–870.
[8] M. Kotani, T. Shirai and T. Sunada: Asymptotic behavior of the transition probabilityof a random walk on an infinite graph, J. Funct. Anal. 159 (1998), pp. 664–689.
[9] T. Sunada: Topological Crystallography: With a View Towards Discrete GeometricAnalysis, Surveys and Tutorials in the Applied Mathematical Sciences 6, SpringerJapan, 2013.
[10] B. Trojan: Long time behaviour of random walks on the integer lattice, available atarXiv:1512.09035.
Ryuya NambaGraduate School of Natural Sciences,
Okayama UniversityOkayama, 700-8530 Japan
e-mail address: [email protected]
(Received January 31, 2017 )(Accepted April 11, 2017 )